“Tomographic” imaging techniques make images of multiple slices of an object. Multiple tomographic images can then be aligned and assembled using a computer to provide a three dimensional view. Some commonly used tomographic imaging techniques include magnetic resonance imaging (MRI) and magnetic resonance spectroscopy (MRS) techniques, which are ideal for assessing the structure, physiology, chemistry and function of the human brain and other organs, in vivo. Because the object of interest is often imaged in many slices and scanning steps in order to build a complete three dimensional view, scans are of long duration, usually lasting several minutes. To increase resolution (detail) of a tomographic scan, more slices and more scanning steps must be used, which further increases the duration of a scan. Magnetic resonance and other long duration imaging techniques (including tomographic techniques), now known or hereafter invented (hereinafter collectively referred to as “MR” or “MRI”) can also afford relatively high spatial and temporal resolution, are non-invasive and repeatable, and may be performed in children and infants.
In addition to MR, other types of scans require multiple repeated exposures, separated in time, of an entire (not slices) object (such as an organ), such as angiograms, in which a dye is injected into a blood vessel and then scans separated in time are taken to determine how and where the dye spreads. These types of scans that detect motion inside a patient or other object over time (“digital angiography systems”) can also have a long duration, and be subject to the problem of patient or object motion.
Many tomographic imaging techniques rely on detecting very small percentage changes in a particular type of signal, which makes these techniques even more susceptible to movements. In functional magnetic resonance imaging, for example, changes in the properties of blood in brain areas activated while subjects are performing tasks causes small signal changes (on the order of a few percent) that can be detected with MR. However, these small signal changes may easily be obscured by signal changes of similar or even greater size that occur during unintentional subject movements.
Because tomographic techniques require that so many images be taken (because so many slices and scanning steps are necessary), the scan has a long duration, so that motion of the subject is a substantial problem for acquiring accurate data. Consequently, subjects commonly are required to lie still to within one millimeter and one degree over extended time periods. Similar requirements exist for other modern imaging techniques, such as Positron Emission Tomography (PET), Single Photon Emission Computerized Tomography (SPECT) and “computer tomography” (CT). These strict requirements cannot be met by many subjects in special populations, such as children and infants, very sick patients, subjects who are agitated perhaps due to anxiety or drug use, or patients with movement disorders, resulting in data with motion artifacts. Similarly, it is exceedingly difficult to perform scans in awake animals.
The basic problem is that it may take several minutes for a scan to be completed, but the patient or other object being scanned cannot remain still for several minutes. Further, the space for a patient or other object being scanned (the “scanning volume”) in an MR machine is very limited—there is very little space in an MR machine once a patient has been positioned inside for a scan.
Several techniques have been developed over the past decades to reduce the sensitivity of scans to motion of the patient or other object being scanned.
Early techniques utilized specially designed scan sequences (“first-order flow/motion compensation”) to minimize the effects of motion. While these approaches are particularly useful for reducing artifacts (or imaging errors) due to flowing blood, swallowing or eye movements, they afford little improvement during movements of entire organs, such as head movements.
Articles entitled “Self-navigated spiral fMRI: interleaved versus single-shot” by Glover G H, et al, in Magnetic Resonance in Medicine 39: 361-368 (1998), and “PROPELLER MRI: clinical testing of a novel technique for quantification and compensation of head motion” by Forbes K, et al, in the Journal of Magnetic Resonance Imaging 14(3): 215-222 (2001), both incorporated herein by reference, disclose how improved sampling schemes for the MRI data can reduce sensitivity to motion. These techniques can reduce motion sensitivity of MR scans under certain conditions, but cannot eliminate errors from motion under all conditions or for very quick movements.
With certain modern ultra-fast “single-shot” imaging techniques (such as “echo-planar imaging”), the entire head (or other organ of interest) is scanned continuously every few seconds (over the course of minutes), for instance, for “functional MRI”. This makes it possible to determine the “pose”, defined as position and rotation, of the head at each instant relative to the initial pose, using image registration (alignment of images). Once the pose for a given instant is known (relative to the initial image), the scanner's image for that instant can be re-aligned to the initial image. For example, the article entitled “Processing strategies for time-course data sets in functional MRI of the human brain” by Bandettini P A, et al, in Magnetic Resonance Medicine 30: 161-173 (1993), incorporated herein by reference, disclosed how realignment of MRI volumes (consisting of multiple slices) can be used to correct for head motion in functional MRI time series. However, these methods are inherently slow because they use MRI, i.e. they correct movements only every few seconds, and are unable to correct for motion in certain directions (orthogonal to the scan planes; in other words, towards or away from the planes in which the scans are being taken).
While all of these techniques reduce sensitivity to subject motion, several problems remain. One major problem is related to the manner in which typical tomographic imaging methods acquire data. Specifically, the data for each cross section (slice) is acquired by moving step by step along “lines” in a mathematical space (“k-space”). The data acquisition step is typically repeated hundreds of times until all lines in the k-space have been filled. For all methods described above, even if motion sensitivity for each individual acquisition (defining a line in k-space) is reduced, these methods typically do not account for variations in head pose amongst the different k-space lines. Second, the methods poorly tolerate fast movements within individual acquisition steps. Finally, one of the most significant issues is that none of these techniques can be applied universally across all the various scanning methods (pulse sequences—the order and manner in which slices are imaged) used in MRI or other tomographic scanning techniques.
One of the most promising approaches to motion correction is to track the pose of the head, brain or other organ of interest (or other object) in real time, during a scan, and to use this pose information to compensate for the detected motion in data acquisitions for subsequent slices within the same scan. This is called adaptive imaging, because the image is adapted during the scan to compensate for the detected motion.
One important aspect of adaptive imaging is the accuracy (or “resolution”) of the motion tracking system. Because of the high resolution needed for medical imaging, the motion tracking system must also have a high resolution, because the motion tracking system's information will be used to align the images of each slice. If the motion tracking system's resolution is high enough, each of the scan images can be accurately aligned (registered) despite a patient's motion.
An article entitled “Prospective multiaxial motion correction for fMRI” by Ward HA, et al, in Magnetic Resonance in Medicine 43:459-469 (2000), incorporated herein by reference, discloses the use of “navigator” signals to estimate the pose of the head and to dynamically correct for head motion.
An article entitled “Spherical navigator echoes for full 3D rigid body motion measurement in MRI” by Welch E B, et al, in Magnetic Resonance in Medicine 47:32-41 (2002), incorporated herein by reference, discloses the use of an MR-based navigator for adaptive motion correction in MRI.
Similarly, an article entitled “Endovascular interventional magnetic resonance imaging.” by Bartels L W, et al, in Physics in Medicine and Biology 48(14): R37-R64 (2003), and another article entitled “Real-time, Interactive MRI for cardiovascular interventions” by McVeigh E R, et al, in Academic Radiology 12(9): 1121-1127 (2005), both of which are incorporated herein by reference, disclose the use of small radiofrequency (RF) coils for tracking catheters during interventional MRI.
While these MR-based “adaptive MRI” techniques provide good results in many situations, they intrinsically interfere with MR acquisitions, work only for a limited number of MR sequences, and are limited to measuring the position or pose a few times per second only.
In order to overcome these shortcomings, recent approaches to real time (“on the fly”) motion correction utilize optical techniques to track subject motion, rather than MR-based methods. The pose information from the tracking system is sent to the scanner and used by the scanner to compensate for the motion in real time. Optical systems are very suitable among alternative tracking technologies because they provide accurate, non-contact sensing with a passive and non-magnetic target. In particular, stereovision (SV) systems have been used for motion tracking for medical imaging.
Stereovision systems employ a target with 3 or more visible landmarks, and at least 2 tracking cameras. By detecting the landmarks in images captured by the cameras and comparing their measured positions and shapes to the known shape of the target, the target position and orientation can be determined. SV systems offer important features including sub-millimeter accuracy when fully calibrated, and update rates limited only by the camera and computing hardware.
However, SV systems have three limitations for adaptive MR imaging: (1) measurement accuracy decreases as the distance between the cameras becomes smaller, (2) the accuracy of orientation measurement decreases as the target becomes smaller; and (3) SV systems have high sensitivity to errors in internal calibration, i.e. small errors in the relative position or rotation of the cameras may cause large errors in the measured target pose. Therefore, SV systems require periodic recalibration. However, accurate calibration has to be performed manually, using a specialized calibration tool or target, is time consuming, and cannot be done while patients are being scanned.
Furthermore, stereovision systems achieve their best accuracy when the separation distance between the cameras is comparable to the distance between the cameras and the target. However, this ideal separation is not possible in an MR scanner because the opening to the scanning volume (the volume which can be scanned by the scanner) is relatively narrow, making it impossible to move the cameras sufficiently far apart and still view into the scanning volume. Additionally, tracking with SV cameras works optimally with larger tracking targets; however, the space in the MR or other scanner environment is very limited.
As noted above, slight errors in the internal calibration of SV systems can produce large measurement errors. For example, an article entitled “Prospective Real-Time Slice-by-Slice 3D Motion Correction for EPI Using an External Optical Motion Tracking System” by
Zaitsev, M C et al, ISMRM 12, Kyoto (2004), which is incorporated herein by reference, tested the use of an SV system for adaptive functional MRI. The system was able to provide 0.4 mm accuracy when ideally calibrated. However, the study contains information showing that a tiny 1/100th degree change in the camera alignments can produce a 2.0 mm error in the position measurement and the study co-authors privately communicated to the present inventors that maintaining calibration was impracticably difficult. Even with extremely careful and rigid engineering of the camera module of an SV system, a measurement drift on the order of 1 mm can be observed while the SV motion tracker warms up, and recommend warm-up periods are 1 to 1.5 hours to avoid drift. Tremblay M, Tam F, Graham S J. Retrospective Coregistration of Functional Magnetic Resonance Imaging Data Using External Monitoring. Magnetic Resonance in Medicine 2005; 53:141-149, incorporated herein by reference.
The prior art has no means to track or correct for these slow changes while the medical imaging system is in service, imaging patients. The error which accumulates in the co-registration, because of loss of camera calibration, is a severe problem for motion compensation in medical imaging using an external tracking system.
As a result, an SV tracking system requires frequent recalibration to accurately determine its position relative to the imaging system. The recalibration procedure involves scanning a specialized calibration tool or sample (“phantom”) at multiple, manually-adjusted positions, both with the Medical imaging system and the SV system. An article entitled “Closed-form solution of absolute orientation using unit quaternions” by Horn, B K P, J. Opt. Soc. Am. 1987; 4:629-642, which is incorporated herein by reference, describes the commonly used “absolute orientation” method. However, since time on a medical imaging system is limited and expensive, removing patients and conducting repeated recalibration with a specialized calibration tool is prohibitively expensive.
Furthermore, Zaitsev et al utilized a relatively large reflective marker, approximately 10 cm (4 inches) in size, which was affixed to the subjects' head in the scanner by means of a bite bar. While a bite bar may be tolerated by healthy and cooperative volunteers, it is an impractical solution for sick or demented patients, or young children.
Therefore, while stereovision systems are able to track subject motion for use with adaptive imaging techniques when conditions are ideal, the use of SV systems for routine clinical scans proves impractical due to cumbersome recalibration procedures, instabilities over time, and awkward size and attachment of tracking markers (i.e. large marker requiring use of a bite bar).
Motion tracking can be improved using prediction means to predict motion, including (without limitation) motion filter and prediction methods. For adaptive MR imaging, the scanner controller requires values of the subject pose at the exact instant adjustments to the scan are applied (Scanning Timing Information). The determination of the subject pose based on actual measurements is an estimation problem. The simplest estimator takes the most recent measurement as the current pose. This simple estimator has been used frequently, for example in an article entitled “Prospective Real-Time Slice-by-Slice 3D Motion Correction for EPI Using an External Optical Motion Tracking System” by Zaitsev, M. C., et al, ISMRM 12, Kyoto (2004), incorporated herein by reference.
However, this simple estimator neglects three-types of information that can improve the accuracy of the estimate of subject pose: (1) measurements prior to the most recent measurement may add information (reduce the covariance of the estimate) if those prior measurements disclose a velocity of the subject's motion; (2) a biomechanical model, in conjunction with the measurement statistics, can be used to constrain the estimated motion (the subject's body only moves in certain ways); and (3) information about the lag time between the pose measurement and the time of the MR scans. By utilizing these additional sources of information, the accuracy of motion tracking and thus of adaptive imaging will be enhanced.
Extended Kalman filtering, which is essentially model-based filtering with simultaneous estimation of the signals and their statistics, is statistically optimal in certain cases and is the most effective framework for incorporating information of types (1), (2) and (3). Kalman filtering has a long history of use in aerospace applications, such as target tracking, aircraft guidance and formation flying of spacecraft, for example in U.S. Pat. No. 5,886,257 “Autonomous Local Vertical Determination Apparatus and Methods for a Ballistic Body,” incorporated herein by reference, which teaches the use of Kalman filtering applied to inertial signals. Kalman filtering has also been previously demonstrated for head motion tracking, for example in “Predictive Head Movement Tracking Using a Kalman Filter”, IEEE Trans. on Systems, Man, and Cybernetics Part B: Cybernetics 1997; 27:326-331, by Kiruluta A, Eizenman M, and Pasupathy S, incorporated herein by reference. Kalman filtering is also disclosed in U.S. Pat. No. 6,484,131 entitled “Localization and Tracking System”, incorporated herein by reference.
Of course, persons of ordinary skill in the art are aware that the prediction means can be implemented in hardware, software, or by other means, and that there are equivalent processes and algorithms to perform the prediction function of the motion filtering and prediction means disclosed above.
U.S. Pat. Nos. 5,936,722, 5,936,723 and 6,384,908 by Brian S. R. Armstrong and Karl B. Schmidt, et al, which are incorporated herein by reference, disclose “Retro-Grate Reflectors”, or RGRs, which allow accurate and fast position measurements with a single camera and a single, relatively small and light orientation marker. The RGR allows the visual determination of orientation with respect to the six degrees of freedom (the three linear directions of left and right, up and down, and forward and back, plus the three rotational directions of roll (rotation around a horizontal axis that points straight ahead), pitch (rotation around a horizontal axis that points side to side) and yaw (rotation around a vertical axis that points up and down)) by viewing a single marker. Pose (position and rotation) is orientation with respect to the six degrees of freedom. As used herein, an object orientation marker is any marker, such as an RGR marker, from which at least three degrees of freedom can be determined by viewing or otherwise remotely detecting the marker.